Basis Talk: Reasoning via Semantic Transformations
summary
Describes a working system Cairo (built on Pyro) that runs SIR ODE models, performs Bayesian posterior inference (variational first, MCMC supported), and expresses causal interventions via effect handlers to simulate lockdowns and noisy interventions. Proposes parametric inversion: add a parameter θ to invert non‑invertible programs, composes inverses to produce sound and complete parametric inverses but admits partiality where some θ values are invalid. Notes implementation issues: twin‑world redundancy, need for compiler optimizations, staging and semantic modifiers; an organization called Basis is building this stack.
AI-written summary · about